Ben Thompson on Big Tech, China, and the AI Boom Running Out of Money - [Invest Like the Best, EP.487]

Summary

• Thompson argues a decisive US “win” in AI would be dangerous, because in a fantastical scenario of American AI-driven military superiority, China’s game-theory-optimal response is “to blow up TSMC.” He calls the reshoring counter “magical thinking”: dependence on China across fabs, actuators, and precursors “is not going to be fixed outside of a conflict,” and even Apple’s India move is diversification, not exit — insurance nobody pays when it’s astronomically expensive. He also doubts AI will necessarily fix real-world manufacturing.
• The current equilibrium — OpenAI and Anthropic on the frontier, Chinese labs distilling to stay “about six to nine months behind” — is “generally favorable to the US,” and may hold longer than people think. The open-source “free” narrative is “bizarre to me”: Kimi is “very expensive to serve” — you skip the R&D, not the inference.
• The nearest-term constraint on the buildout may be neither compute nor power but money. The buildout started with free cash flow, then tech “blew through the debt markets” in about a year; Google is issuing equity, and NVIDIA is assembling a $500 billion vehicle to tap pension funds and insurance floats — “What’s after that?” The railroad precedent: in the 1870s “the world just ran out of money,” yet the railroads kept operating and expanded the West.
• Google’s shocking equity issuance raises the possibility that Berkshire is a model for Google: Search could be See’s Candies, and AI could be BNSF. Past a certain capital scale you operate on absolute rather than percentage numbers — lower-margin AI against a TAM of “basically all white-collar work” can dwarf perfect-margin search, and a smaller share of “an astronomically larger pie” leaves no one complaining.
• Tech in general doesn’t understand commodity markets, and compute is running a classic cycle: today’s money manifests as compute in 2028-29, while payback periods are “measured in a time of scarcity.” Jassy’s and Nadella’s we-only-buy-GPUs-on-demand story draws a hedge — “I’m not sure that I think it’s a lot of BS” — because sunk shells may be filled, and even if AI demand is endless, an “air gap” blowup is possible if revenue arrives after capital runs out.
• TSMC’s conservatism offloaded risk onto big tech as foregone profit — and that scarcity “is what ultimately saved Intel.” Nobody rationally pays negative-expected-value insurance, so the only fix for TSMC dependence was a compute use case so massive that big tech is economically forced to bring Intel and Samsung up to speed; Thompson expects a major Intel partner announcement.
• NVIDIA’s margins are “unnatural”: the 25% backstops and circular financing are price cuts in disguise, because “risk never disappears — it just moves.” Its real competitors are hyperscalers with lower cost of capital selling TPUs (about 20% to Anthropic) and Trainium externally as commodities; the longer power remains sufficient, the more time those chips have to catch up. The bubble’s lasting legacy could be energy abundance — AI’s version of WorldCom fiber.
• OpenAI replayed the Dropbox mistake “at like a hundred X the size”: consumers won’t pay for software and don’t want to be productive. Had it leaned into ads the moment ChatGPT hit, “Google would be in much bigger trouble… Meta would be in much bigger trouble” — instead, the biggest LLM monetization today is probably the incremental ad gain at Google and Meta, whose marketplaces are verification machines at global scale.

Deep dive

1. A decisive US AI “win” could invite China to blow up TSMC

• Thompson’s opening provocation: “I think it would be very problematic for the US to win.” In the most fantastical scenario, where controlling AI confers military superiority and even fixes manufacturing — something he doubts AI will necessarily do because it involves the real world — “what is the game theory optimal response of China? To blow up TSMC.” He describes a “fundamental disconnect” with the rhetoric of “one of the labs in particular” that treats national-security dominance as the goal rather than the danger.
• Patrick’s pushback — wouldn’t the US have domestic fabs by then? Thompson: “a little bit of magical thinking.” Dependence on China across fabs, actuators, and precursors “is not going to be fixed outside of a conflict,” because reshoring against a China-sourcing competitor is ruinous; even Apple’s India shift is partial — an insurance policy you skip when it’s astronomically expensive.
• On motivations: “Everyone can use a good bogeyman… from the AI trade perspective, nothing works better than we have to beat China.” His prescription is anti-mimicry: “more openness, more innovation, less top-down control” — “America succeeds by being on the leading edge and by leaning into that.”

2. Today’s equilibrium favors the US — and “free” open source isn’t free

• The AI race resembles Taiwan itself: a status quo that “actually doesn’t seem so bad,” of unknown duration. OpenAI and Anthropic are clearly on the frontier, “who knows what’s happening with Google,” and Groq and Meta are chasing. Chinese labs are distilling to stay six to nine months behind. He’s “still a little skeptical” that China pulls ahead — chips are one reason, and it is difficult to catch a frontier accelerating via AI-improving-AI, which “seems to be coming true” at both leading labs.
• On open models: “Everyone referring to these as free… No, you can’t use AI for free.” GLM and Kimi carry real inference costs — “Kimi is very expensive to serve” — and if frontier labs use AI to optimize their own stacks, their marginal cost to serve may end up structurally lower too.
• His biggest known-unknown: after the scare around “Mythos” and the “Hugging Face incident,” labs may respond not by reducing dangers but by not releasing — everyone judging Mythos by Fable while losing “any sense of what is actually the frontier,” a false security whose gap “is only going to increase over time.”

3. The boom may run into a funding gap — and Google’s raise reveals a Berkshire analogy

• A mismatch that worries him is capital timing: “We’re working our way down the capital curve.” The buildout started with free cash flow; then “the speed with which the tech companies blew through the debt markets is kind of incredible — it took, like, a year”; now Google is issuing equity and NVIDIA’s $500 billion vehicle is tapping pension funds and insurance floats. “What’s after that?” If cash flow doesn’t flip back in time, “we could have a big blowup” — against Patrick’s backdrop of ~$800B CapEx this year and an estimated $1.3T next.
• The railroads were exactly this: decade-long payoffs funded short-term until, in the 1870s, “the world just ran out of money.” Yet the railroads kept operating and expanded the West — and, deliciously, “railroad money is what’s going into Google right now” via Berkshire. Even a blowup would not stop AI from improving; Thompson believes its eventual economic impact will be astronomical and its societal effects significant. Bubbles are “ultimately immaterial in terms of the broad scope of humanity, even if they were very devastating.”
• Why Google’s equity issuance shocked him: “It’s Google… Why are they reducing their upside if they believe so strongly in this?” His answer runs through See’s Candies and BNSF — one year of BNSF free cash exceeded See’s lifetime output — because at sufficient capital scale “you start operating in a world of absolute numbers as opposed to percentage numbers.”
• His proposed mapping: Search is “one of the most perfect, beautiful business models of all time, and the purest aggregator of them all,” while AI is “just incinerating cash” — but if AI’s TAM is “basically all white-collar work” and eventually, with robotics, everything, a diluted share of “an astronomically larger pie” leaves nobody complaining. Berkshire may be “not just an investor in Google, but a model for Google.”

4. Bullish and unconvinced at once: the verifiable-domain problem

• His honest split — “super bullish and less bullish in some respects.” AI is incredible at coding and math, but those are verifiable domains, and he wants the example that isn’t chess or Go: he was “kind of annoyed” when lab panelists took him for a bear, because “I thought we could solve chess, I thought we could solve Go — they’re knowable domains.” Where is AI succeeding in “genuinely… an unknowable space”?
• One resolution hypothesis: models distilled “the end state of human thought” — the Reddit comment, not the traces of thought and emotion behind it. “What if the actual payoff from Neuralink is actually capturing the traces of human thought” as training data that dramatically expands capability?
• Even frozen at today’s capability, the market is massive: “there’s a lot of people in the world who are kind of like sentient AIs” — operating well in verifiable domains, given jobs, doing them. Hence his label, “reluctant accelerationist”: “we can’t go back, and the worst thing we can do is get stuck where we are.”
• Medicine is “by far one of the biggest opportunities” — machine learning on all medical records would yield discoveries “in a very rapid amount of time” — but regulation gates it. The double-edged coda: human “capacity to create needs is sort of unlimited,” yet “our ability to create red tape and muck is also fairly unlimited.”

5. Marginal costs are back — and they break Microsoft’s pricing model

• Aggregation theory rested on zero marginal cost: “in a world of abundance, the hard problem is not distribution, it’s discovery,” and monetization scaled human-free — most Google and Meta advertisers never talk to a person. Patrick’s pushback that inference costs “are real right now” draws a counter: the spread is the point. The recipe-maker ChatGPT user costs roughly a webpage to serve, while test-time scaling is “directly marginal cost — every second longer you’re thinking is costing more money.” The two users are “not even remotely in the same universe.”
• Microsoft’s new E7 plan — $100 per user per month plus usage charges — is “fraught”: companies budget annually, not monthly; per-seat software was thoughtlessly baked into headcount, “not CapEx, but kind of like CapEx”; and monthly bill scrutiny invites “what am I paying for? How good is each of these products?” They had to do it — heavy token users cost far more than $100 — but they need customers to think a little “but not too much, because that breaks the model.”

6. Consumers won’t pay and won’t be productive — OpenAI replayed Dropbox at 100x

• Two things “Silicon Valley has to relearn every ten years: consumers do not wanna pay for software, and consumers do not care about being productive.” The canonical case is Dropbox: Jobs’ “you’re a feature, not a company,” a two-year lull, and a ground-up rebuild for enterprise permissions — because companies pay when employees get more productive. Consumers: “I spent all day working. Why do I wanna come home and be more productive? I wanna sit on the couch and watch Reels.”
• “You literally had OpenAI replaying the Dropbox story but at like a hundred X the size.” Subscriptions sold a lot “but they didn’t sell enough,” and the ad pivot lands awkwardly just as “Anthropic is kicking our rear end” pulls them toward enterprise. Had they leaned into ads immediately: “I think that Google would be in much bigger trouble. I think Meta would be in much bigger trouble.”
• Why advertising wins with consumers: “your ability to monetize the consumer is infinite because the advertiser is bearing the price increase” — no Netflix-style elasticity wall. “Charging people money is hard. Giving people things for free is easy.”

7. Tech doesn’t understand commodity markets: today’s money is 2028’s compute

• The shortage exists because of under-investment in 2023-24, and TSMC decreased its growth rate in 2023, 2024, and 2025 — so “our shortage of compute is going to get worse,” and today’s spending “all manifests in compute in 2028 and 2029.” Jassy’s and Nadella’s we-only-buy-GPUs-when-demand-appears story draws a hedge: “I’m not sure that I think it’s a lot of BS” — a built shell is sunk capital you won’t leave idle.
• His shipping tutorial: a ship’s cost is depreciation, “an accounting figment — you already paid the money,” so you run it regardless and price falls toward marginal cost. COVID took containers from $3,000-4,000 to $17,000-18,000; if everyone orders ships on two-year lead times, prices can plummet. Suppliers exit only when prices fall below real marginal cost, reducing supply and allowing prices to recover. Memory’s boom-bust history is the same mechanism.
• The trap for data-center economics: “you’re measuring your payback period in a time of scarcity. Is that payback period gonna hold in a time of abundance?” Bulls say test-time scaling means “we’re gonna be short forever” — “maybe we will be” — but even then “we could still have an air gap” if revenue arrives after the capital does.

8. Memory made itself a target; TSMC’s caution is what saves Intel

• Memory consolidated to a disciplined three after brutal cycles — Samsung “invested into downturns,” which took “a ton of guts and a ton of discipline,” and “basically wiped out the Japanese” — and that discipline collided with a secular demand shift they were slow to recognize. His analogy: memory makers are Iran, and using the Strait of Hormuz invites being built around — “say Iran wants to close the Strait of Hormuz in 2035, it’s not gonna have any effect.” The number-one algorithmic focus now: use less memory.
• TSMC is “arguably worse because there’s only one.” Its conservatism (a fab must run thirty years) didn’t eliminate risk — “risk doesn’t disappear. It just moves” — it moved onto big tech as foregone revenue and profits. Morris Chang came out of retirement after new leadership cut spending during the Great Recession, fired everyone, and said: “The iPhone just launched… We need to be investing, not cutting.” Chang is “a one of one on the Mount Rushmore” of tech executives.
• It never made rational sense to endure the pain of training Intel when TSMC is “so great to work with” — until the shortage got acute enough. “The scarcity is what ultimately saved Intel… ultimately, TSMC brought it on themselves”; he expects a major Intel partner announcement. The general law: nobody pays negative-EV insurance, so the fix was a compute use case so massive “we get the geopolitical insurance for free.” Patrick’s gloss: “the cure for high prices is high prices.”

9. Amazon is the setup he likes most; Apple may be better be lucky than good

• Asked for the best big-tech setup: “The answer’s always Amazon” — they are their own “first best customer.” AWS wasn’t spare capacity (it served external customers before internal ones); logistics ran the reverse path; early Graviton and Trainium were “terrible,” but hidden under managed services like Redshift they got volume to improve, and Trainium now runs Anthropic. Core retail “feels so impervious” — “their moat feels deeper than anyone.”
• Apple sitting out AI “might be a situation of better be lucky than good”: “if you own access to customers, suppliers come to you,” and a consumer chatbot could eventually run on-device — “using the customer’s electricity,” no inference bill. The risk is the Microsoft-mobile trap: assuming the phone stays the center while ambient AI manifests everywhere.
• The temperamental point: “AI is this probabilistic endeavor. Apple is the king of deterministic products” — never an iPhone recall. “I’m fine with Apple not doing AI. I want them to keep making great devices.”

10. “They think they’re creating God” — and Microsoft runs the IBM playbook

• On the frontier labs: “Never discount the power of belief. They think they’re creating God.” OpenAI is “kinda like mainline — they go to church every Sunday”; Anthropic are the evangelicals, “all in.” Meta on the frontier is “one of the purest manifestations of founder energy, for better or for worse”; Google just needs search “to not die too quickly.” The space-data-center play (“SpaceX AI,” as spoken) has the weakest case for owning a model — if orbital capacity is the differentiation, “why are you wasting billions?”
• Microsoft is deliberately off the frontier — hence $40 billion of quarterly free cash flow and a $10 billion dividend — running Gerstner’s IBM play: mediocre at everything (“that’s the price of monopoly”), so sell middleware, dependability, and consulting, the way IBM “brought all of corporate America online.” All enterprise sales is “companies whose long-term goal is to lock you in getting you on board by trying to make you scared of being locked into somebody else” — the Oracle joke.
• Rational and desperate at once: “at the end of the day, why are we using Microsoft products again?” AI is “actually surprisingly good” at the tedious migrations that protect systems of record, and Codex/Claude-style agents are “aimed like an arrow to the heart” of the interface layer Microsoft owns. Hence the broader claim: for digital companies “it is more reckless to not be on the frontier.”

11. Meta’s ad marketplace is a verification machine at global scale

• Meta’s structural quirk: Instagram “generates all this money for which Facebook pays zero dollars for content,” so AI-generated content, if Meta is generating it, has a worse margin profile for Meta while potentially being accretive for YouTube, whose inference could cost less than creator revenue share. The bull case: in an AI-saturated world “the desire for a human connection becomes greater” — social networking, which TikTok showed was an artificial constraint on content quality, might matter again.
• The most under-told point: “the biggest monetization right now is probably not Anthropic or OpenAI. It’s the incremental gain that is happening for Google and Meta.” Their ad marketplaces verify generated creative through global-scale human click and purchase verdicts and extensive A/B testing; most ads are throwaway, but LLM-style prediction — “this person probably wants to see this next” — needs only a few percentage points of improvement “for the returns to be billions and billions of dollars.” That alone justifies frontier investment.
• His frustration: Zuckerberg “has never really talked about the societal benefits of advertising except in passing in twenty years,” and the Sheryl Sandberg case-study evangelism “never got filled.” Meanwhile, Thompson calls Apple’s ATT “one of the worst antitrust violations in the history of technology” and describes its bus advertisement as a “dishonest” representation — and Wall Street’s reluctance to fund Meta’s AI spend traces partly to the “cumulative hundred-some billion dollars on Oculus, which I hated all along.”

12. NVIDIA’s margins are unnatural — and power abundance could be the real legacy

• Patrick’s framing question — are compute and intelligence both commodities? — gets an embrace, not a dodge: “Commodities change the world… The internet is a commodity. It changed the world, so I don’t think it’d be weird that intelligence ends up a commodity and changes the world.” Differentiated products cap their own TAM; commodities are impactful precisely because everyone can have them.
• NVIDIA’s headline margins are being maintained by moving price cuts off the P&L: the 25% backstop, equity in neoclouds, and compute-purchase guarantees to 2030 lower counterparties’ cost of capital “because Nvidia assumed risk… Risk never disappears. It just appears somewhere else.” Priced honestly, that expected-value transfer “is a price cut” — “we have seen price cuts. They’re just manifesting in these very bizarre sort of ways.”
• The real competitors are the hyperscalers: Google selling roughly 20% of its TPUs to Anthropic, Jassy practically confirming external Trainium sales — sold as commodities, not differentiation, backed by “a lower cost of capital. It’s a capital fight.” CUDA’s moat is “dramatically diminished because the models don’t care what they run on”; Elon buys NVIDIA not “because they’re the best” but “‘cause they’re the most fungible.”
• Thompson thinks insufficient power may have been expected to become NVIDIA’s moat sooner — token efficiency wins in a constrained world — but the US “brought a lot more power online than expected” (behind-the-meter builds, West Texas gas, nuclear restarts), buying Trainium and TPUs time; “those margins seem very hard to sustain.” And the bubble’s lasting artifact could be energy itself, the way “our core internet still runs on WorldCom fiber”: “What would it be like to live in a world of energy abundance? It’s hard to even imagine.”